A novel approach to axiomatic decision theory

post by lukeprog · 2012-05-26T21:18:30.111Z · LW · GW · Legacy · 21 comments

Contents

21 comments

In the standard approach to axiomatic Bayesian decision theory, an agent (a decision maker) doesn't prefer Act #1 to Act #2 because the expected utility of Act #1 exceeds that of Act #2. Instead, the agent states its preferences over a set of risky acts, and if these stated preferences are consistent with a certain set of axioms (e.g. the VNM axioms, or the Savage axioms), it can be proven that the agent's decisions can be described as if the agent were assigning probabilities and utilities to outcomes and then maximizing expected utility. (Let's call this the ex post approach.)

Peterson (2004) introduces a different approach, which he calls the ex ante approach. In many ways, this approach is more intuitive. The agent assigns probabilities and utilities directly to outcomes (not acts), and these assignments are used to generate preferences over acts. Using this approach, Peterson claims to have shown that the principle of expected utility maximization can be derived from just four axioms.

As Peterson (2009:75,77) explains:

The aim of the axiomatization [in the ex ante approach] is to show that the utility of an act equals the expected utility of its outcomes.

...The axioms... entail that the utility of an act equals the expected utility of its outcomes. Or, put in slightly different words, the act that has the highest utility (is most attractive) will also have the highest expected utility, and vice versa. This appears to be a strong reason for letting the expected utility principle guide one's choices in decision under risk.

Jensen (2012:428) calls the ex ante approach "controversial," but I can't find any actual published rebuttals to Peterson (2004), so maybe Jensen just means that Peterson's result is "new and not yet percolated to the broad community."

Peterson (2008) explores the ex ante approach in more detail, under the unfortunate title of "non-Bayesian decision theory." (No, Peterson doesn't reject Bayesianism.) Cozic (2011) is a review of Peterson (2008) that may offer the quickest entry point into the subject of ex ante axiomatic decision theory.

Peterson (2009:210) illustrates the controversy nicely:

...even if [this] discussion may appear a bit theoretical... the controversy over [ex post and ex ante approaches] is likely to have important practical implications. For example, a forty-year-old woman seeking advice about whether to, say, divorce her husband, is likely to get very different answers from the [two approaches]. The [ex post approach] will advise the woman to first figure out what her preferences are over a very large set of risky acts, including the one she is thinking about performing, and then just make sure that all preferences are consistent with certain structural requirements. Then, as long as none of the structural requirements is violated, the woman is free to do whatever she likes, no matter what her beliefs and desires actually are... The [ex ante approach] will [instead] advise the woman to first assign numerical utilities and probabilities to her desires and beliefs, and then aggregate them into a decision by applying the principle of maximizing expected utility.

I'm not a decision theory expert, so I'd be very curious to hear what LW's decision theorists think of the axiomatization in Peterson (2004) — whether it works, and how significant it is.

21 comments

Comments sorted by top scores.

comment by lukeprog · 2012-05-26T22:14:31.227Z · LW(p) · GW(p)

Off-topic, but I cannot resist sharing Peterson's story of how he became interested in decision theory:

A while back a beautiful woman, whom I quite liked, asked me to marry her. I was stunned. Marriage? Now? It is too early! I have not even turned forty! However, for one reason or another I decided not to share my spontaneous reaction with her. I said I felt overwhelmed by this… unexpected proposal, and that I need some time ot think it over. At dawn the following day I… raced to the university library. I borrowed all the books I could find on decision theory. Later the same afternoon, after having learned what modern decision theory is all about, I still had no clue how to answer the lady.

comment by Wei Dai (Wei_Dai) · 2012-05-31T22:03:42.645Z · LW(p) · GW(p)

I spent quite a few hours going through Peterson's 2008 book (online copy available here) to see if there were any interesting ideas, and found the time largely wasted. (This was my initial intuition, but I thought I'd take a closer look since Luke emailed me directly to ask me to comment.) It would take even more time to write up a good critique, so I'll just point out the most glaring problem: Peterson's proposal for how to derive a utility function from one's subjective uncertainty about one's own choices, as illustrated in this example:

This means that if the probability that you choose salmon is 2/3, and the probability that you choose tuna is 1/3, then your utility of salmon is twice as high as that of tuna.

What if we apply this idea to the choice between $20 and $30?

Let us now return to the problem of perfect discrimination mentioned above. As explained by Luce, the problem is that ‘the [utility] scale is defined only over a set having no pairwise perfect discriminations, which is probably only a small portion of any dimension we might wish to scale That is, the problem lies in the assumption that p(x  > y) != 0,1 for all x,y in B. After all, this condition is rather unlikely to be satisfied, because most agents know for sure that they prefer $40 to $20, and $30 to $20, etc.

Peterson tries to solve this problem in section 5.3, but his solution makes no sense. From page 90:

Suppose, for example, that I wish to determine my utility of $20, $30, and $40, respectively. In this case, the non-perfect object can be a photo of my beloved cat Carla, who died when I was fourteen. If offered a choice between $20 and the photo, the probability is 1/4 that I would choose the money; if offered a choice between $30 and the photo, the probability is 2/4 that I would choose the money; and if offered a choice between $40 and the photo, the probability is 3/4 that I would choose the money. This information is sufficient for constructing a single ratio scale for all four objects. Here is how to do it: The point of departure is the three local scales, which have one common element, the photo of Carla. The utility of the photo is the same in all three pairwise choices. Let u(photo) = 1. Then the utility of money is calculated by calibrating the three local scales such that u(photo) = 1 in all of them.

So we end up with u($20)=1/3, u($30)=u(photo)=1, u($40)=3. But this utility function now implies that given a choice between $20 and $30, you'd choose $20 with probability 1/4, and $30 with probability 3/4, contradicting the initial assumption that you'd choose $30 with certainty. I have no idea how Peterson failed to notice this.

Replies from: Tyrrell_McAllister
comment by Tyrrell_McAllister · 2012-06-01T18:52:48.629Z · LW(p) · GW(p)

I've only read your comment, not anything by Peterson, so I'm just asking for clarification on what he claims to do:

In your first quote of him, he claims to derive utilities from a certain kind of subjective probability. But does he also claim to make the converse derivation? That is, does he also claim to derive those same subjective probabilities from utilities as you do in your final paragraph? It's not clear to me that your first quote of him commits him to doing this.

Replies from: Wei_Dai
comment by Wei Dai (Wei_Dai) · 2012-06-02T02:23:00.358Z · LW(p) · GW(p)

That is, does he also claim to derive those same subjective probabilities from utilities as you do in your final paragraph?

No, he doesn't commit to doing this, but taking this defense doesn't really save his idea. Because what if instead of thinking I'd take $30 over $20 with probability 1, I think I'd make that choice with probability 0.99. Now u($30)/u($20) has to be 99, but u($30)=u(photo) and u(photo)/u($20)=3 still hold, so we can no longer obtain a consistent utility function using Peterson's proposal. What sense does it make that we can derive a utility function if the probability of taking $30 over $20 is either 1 or 3/4, but not anything else? As far as I can tell, there is no reason to expect that our actual beliefs about hypothetical choices like these are such that Peterson's proposal can output a utility function from them, and he doesn't address the issue in his book.

Replies from: Tyrrell_McAllister
comment by Tyrrell_McAllister · 2012-06-02T04:23:56.937Z · LW(p) · GW(p)

What sense does it make that we can derive a utility function if the probability of taking $30 over $20 is either 1 or 3/4, but not anything else?

It seems that he should account for the fact that this subjective probability will update. For example, you quoted him as saying

This means that if the probability that you choose salmon is 2/3, and the probability that you choose tuna is 1/3, then your utility of salmon is twice as high as that of tuna.

But once I know that u(salmon)/u(tuna) = 2, I know that I will choose salmon over tuna. I therefore no longer assign the prior subjective probabilities that led me to this utility-ratio. I assign a new posterior subject probability — namely, certainty that I will choose tuna. This new subjective probability can no longer be used to derive the utility-ratio u(salmon)/u(tuna) = 2. I have learned the utility-ratio, but, in doing so, I have destroyed the state of affairs that allowed me to learn it. I might remember how I derived the utility-ratio, but I can no longer re-derive it in the same way. I have, as it were, "burned up" some of my prior subjective uncertainty, so I can't use it any more.

Now suppose that I am so unfortunate as to forget the value of the utility-ratio u(salmon)/u(tuna). However, I still retain the posterior subjective certainty that I choose salmon over tuna. Now how am I going to get that utility-ratio back? I'm going to have to find some other piece of prior subjective uncertainty to "burn". For example, I might notice some prior uncertainty about whether I would choose salmon over $5 and about whether I would choose tuna over $5. Then I could proceed as Peterson describes in the photo example.

So maybe Peterson's proposal can be saved by distinguishing between prior and posterior subjective probabilities for my choices in this way. Prior probabilities would be required to be consistent in the following sense: If

  • my prior odds of choosing A over B are 1:1, and
  • my prior odds of choosing B over C are 3:1,

then

  • my prior odds of choosing A over C have to be 3:1.

Thus, in the photo example, the prior probability of taking $30 over $20 has to be 3/4, given the other probabilities. It's not allowed to be 0.99. But the posterior probability is allowed to be 1, provided that I've already "burned up" some piece of prior subjective uncertainty to arrive at that certainty. In this way, perhaps, it makes sense to say that "the probability of taking $30 over $20 is either 1 or 3/4, but not anything else".

Replies from: Tyrrell_McAllister
comment by Tyrrell_McAllister · 2012-06-02T20:33:19.879Z · LW(p) · GW(p)

I wrote,

Now suppose that I am so unfortunate as to forget the value of the utility-ratio u(salmon)/u(tuna).

But, on reflection, the possibility of forgetting knowledge is probably a can of worms best left unopened. For, one could ask what would happen if I remembered that I was highly confident that I would choose salmon over tuna, but I forgot that I was absolutely certain about this. It would then be hard to see how to avoid inconsistent utility functions, as you describe.

Perhaps it's better to suppose that you've shown, by some unspecified means, that u(salmon) > u(tuna), but that you did so without computing the exact utility-ratio. Then you become certain that you choose salmon over tuna, but you no longer have the prior subjective uncertainty that you need to compute the ratio u(salmon)/u(tuna) directly. That's the kind of case where you might be able to find some other piece of prior subjective uncertainty, as I describe in the above paragraph.

comment by Vaniver · 2012-05-27T18:52:14.177Z · LW(p) · GW(p)

Prediction from reading this post before I delved into the paper: the controversy is going to be about psychology, not decision theory. (After delving into the paper: I'm going to go with 'prediction confirmed.')

So, he uses six axioms. How do they map onto Howard's 5 that I used? It looks like his 0 is essentially "you can state the problem," his 1 and 2 are my choice, his 3 and 4 don't seem to have mirrors, and his 5 is my equivalence. I find it a little worrisome that three of my axioms don't appear to show up in his- probability, order, and substitution- except possibly in 0. They're clearly present in his analysis, but they feel like things that should be axioms instead of just taken for granted, and it's not clear to me why he needed to raise 3 and 4 to the level of axioms.

It's also not clear to me why he puts such emphasis on the independence and "sure-thing" principle. The "sure-thing" principle is widely held to only apply to a certain class of utility functions / "sure things", and there's not a good reason to expect people should or do have those utility functions. (Outcomes, properly understood, are the entire future- and so a game in which I flip a coin and you win or lose $100 can be different from a game in which I flip a coin and give you either $0 or $200, because you're $100 richer in the second game.) Similarly, independence only holds for normatively correct probability functions, and so if you allow normatively incorrect probability functions you have to throw out independence.

(The difference between independence and "sure-thing" is that "sure-thing" applies a transformation to all of the outcomes, which may not be the same transformation to all of the utilities of those outcomes, and independence applies a transformation to all of the probabilities, which will be the same transformation to all of the act utilities for normatively correct probability functions.)

For the controversy, I'll quote him directly:

I am aware of two objections to the proposed ex ante axiomatization. The point of departure of the first objection is the reasonable suspicion that people simply do not have numerical utilities in their heads, or at least have no access to them. Therefore, to assume that utility is an extensive structure is unreasonable since it seems practically impossible for decision makers to state the preferences required for obtaining the utility function. The same holds true for the probability function; where does the exogenously defined probability function come from?

The probability complaint is uninteresting. The "how do we measure utilities?" complaint is serious but a little involved to discuss.

Basically, there are three branches of rationality: epistemic, instrumental, and terminal (I'm not sure I like "terminal rationality" as a name; please suggest alternatives): thinking about uncertainties and probabilities, thinking about actions, and thinking about outcomes.

Peterson's root complaint is that traditional decision theory is silent on the valuable parts of terminal rationality- it only ensures that your elicited preferences are consistent and then uses them. If they're consistent but insane, the expected utility maximization won't throw up any error flags (for example, Danzig's diet optimization which prescribed 500 gallons of vinegar a day), because checking for sanity is not its job.

But pointing that complaint at decision theory seems mistaken, because it's a question of how you build the utility function. The traditional approach he describes in lukeprog's post above (the woman considering divorce) uses casuistry, expecting that people can elicit their preferences well about individual cases and then extrapolate. I think the approach he prefers is deconstruction- isolate the different desires that are relevant to the outcomes in question, construct (potentially nonlinear) tradeoffs between them, then order the outcomes, then figure out the optimal action. The first checks primarily for internal consistency; the second checks primarily for reflective equilibrium. But both can do each, and they can be used as complements instead of substitutes.

TL;DR: There does not appear to be meat to the controversy over axioms, and if there is then Peterson's axioms strike me as worse than Howard's and possibly worse than vNM or Savage. There is meat to the controversy over discovering utility functions, but I don't think the 2004 paper you link is a valuable addition to that controversy. Compare it to the chapter on choice under certainty from Thinking and Deciding.

Replies from: lukeprog, JonathanLivengood
comment by lukeprog · 2012-05-28T02:20:27.236Z · LW(p) · GW(p)

I'm confused by several parts of your reply. I'll select just two of them for discussion. Perhaps you'll have the motivation to try to un-confuse me.

  • I don't understand your points about the axioms he uses. He uses the axioms required to derive the useful results he aimed for, given his approach to formalizing decision problems, and no more than that. Do you reject the plausibility of one of his axioms, or do you disagree that his results follow from his axioms?

  • I don't think Peterson denies the usefulness of traditional axiomatic decision theory for checking the consistency of one's preferences, he's just saying that it would also be nice to have a decision theory that can tell you what you should choose given what you believe and what you value. Indeed, this is what many/most people actually do when trying to make decisions "rationally," but this norm wasn't justified with an axiomatic approach until Peterson (as far as I can tell).

Replies from: Vaniver
comment by Vaniver · 2012-05-28T21:59:10.917Z · LW(p) · GW(p)

He uses the axioms required to derive the useful results he aimed for, given his approach to formalizing decision problems, and no more than that.

Can you give me two examples of useful results he derives from the axioms? That'll help me target my response. (I should note that the commentary in the grandparent is targeted at the 2004 paper in the context of the other things you've quoted on this page; if there's relevant material in one of the other links I probably missed it.)

I don't think Peterson denies the usefulness of traditional axiomatic decision theory for checking the consistency of one's preferences, he's just saying that it would also be nice to have a decision theory that can tell you what you should choose given what you believe and what you value.

Agreed. In this comment I want to differentiate between "decision theory" and a component of it, "expected utility theory" (I didn't differentiate between them in the grandparent). The first studies how to make decisions, and the second studies a particular mathematical technique to isolate the highest scoring of a set of alternative actions. My claim is that expected utility theory is and should be silent on the design of human-appropriate utility functions, but that decision theory should include a component focused on the design of human-appropriate utility functions. That component will be primarily researched by psychologists- what makes humans happy, what do humans want, how do we align those, what common mistakes do humans make, what intuitions do humans have and when are those useful, and so on.

Peterson's axioms look to me like trying to shoehorn human-appropriate utility functions into expected utility theory, which doesn't seem to augment the math of calculating expected utilities or augment the actual design of human-appropriate utility functions. As far as I can tell, that field is too young to profit from an axiomatic approach.

But I said "profit" from axioms and you said "justified" with axioms, and those are different things. It's not clear to me that Peterson's axioms are useful at justifying the use of expected utility theory, and my hesitance hinges on the phrase "given what you believe and what you value" from the parent. That means that's Peterson's decision theory takes your beliefs and values as inputs and outputs decisions- which is exactly what traditional decision theory does, and so they look the same to me (and if they're different, I think it's because Peterson made his worse, not better). The underlying problem as I see it is that beliefs and values are not given, they have to be extracted- and traditional decision theory underestimated the difficulty of that extraction.

(Side note: decision theory underestimating the difficulty and decision theorists underestimating the difficulty are very different things. Indeed, it's likely that decision theorists realized the problem was very hard, and so left it to the reader so they wouldn't have to do it!)

Then the question is how much Peterson 2004 helps its readers extract their beliefs and values. As far as I can tell, there's very little normative or prescriptive content.

Replies from: lukeprog
comment by lukeprog · 2012-05-30T22:05:07.533Z · LW(p) · GW(p)

My claim is that expected utility theory is and should be silent on the design of human-appropriate utility functions, but that decision theory should include a component focused on the design of human-appropriate utility functions.

What do you mean by "the design of human-appropriate utility functions"?

Can you give me two examples of useful results he derives from the axioms? That'll help me target my response.

Actually, let me show you a section of Peterson (2009), which is an updated and (I think) clearer presentation of his axiomatic ex ante approach. It is a bit informal, but is mercifully succinct. (The longer, formal presentation is in Peterson 2008). Here is a PDF I made of the relevant section of Peterson (2009). It's a bit blurry, but it's readable.

Replies from: Vaniver
comment by Vaniver · 2012-05-31T05:35:56.669Z · LW(p) · GW(p)

What do you mean by "the design of human-appropriate utility functions"?

A utility function that accurately reflects the beliefs and values of the human it's designed for. Someone looking for guidance would get assistance in discovering what their beliefs and values about the situation are, rather than just math help and a consistency check. Similarly, someone could accidentally write a utility function that drowns them in vinegar, and it would be nice if the decision-making apparatus noticed and didn't.

That's my interpretation of "he's just saying that it would also be nice to have a decision theory that can tell you what you should choose given what you believe and what you value."

Actually, let me show you a section of Peterson (2009), which is an updated and (I think) clearer presentation of his axiomatic ex ante approach.

This looks like it boils down to "the utility of an act is the weighted sum of the utility of its consequences." It's not clear to me what good formulating it like that does, and I don't like that axiom 4 from the 2009 version looks circular. (You're allowed to adjust the utility of different equiprobable outcomes so long as the total utility of the act is preserved. But, uh, aren't we trying to prove that we can calculate the utility of an act with multiple possible outcome utilities, and haven't we only assumed that it works for acts with only one possible outcome utility?)

Was Thm 4.1 an example of a useful result?

comment by JonathanLivengood · 2012-05-28T18:58:48.038Z · LW(p) · GW(p)

Suggestions for replacing "terminal":

"outcome"; "ultimate" (or "ultima" if you prefer Latin); "intrinsic" (to draw a contrast with "instrumental"); "telikos" (transliteration of the Greek for final, ultimate, terminal, or last); "endpoint"; "consequence"; "effect"; "impact"; "goal."

Do any of those sound better to you?

Edit -- slight change owing to formatting.

Replies from: Vaniver
comment by Vaniver · 2012-05-28T22:01:20.928Z · LW(p) · GW(p)

I like "goal", and think I like "value" even more. Value rationality?

Replies from: JonathanLivengood
comment by JonathanLivengood · 2012-05-28T22:16:37.798Z · LW(p) · GW(p)

I'm not sure about "value" in this context. The term "value" could attach to either acts or outcomes, I think. So, if the goal is to distinguish rationality that cares about acts first from rationality that cares about outcomes first, then "value" doesn't seem to do a very good job. Does that sound right to you, or am I missing something about the distinction that you want to draw?

Replies from: Vaniver
comment by Vaniver · 2012-05-28T22:17:30.231Z · LW(p) · GW(p)

That's convinced me that "goal" is clearer than "value."

comment by lukeprog · 2012-05-27T01:01:31.872Z · LW(p) · GW(p)

Peterson's way of formally representing a decision problem also seems more helpful to me than the ways proposed by Jeffrey and Savage. Peterson (2008) explains:

I shall conceive of a formal representation as an ordered quadruple π =<A,S,P,U>. The intended interpretation of its elements is as follows.

A = {a1,a2, . . .} is a non-empty set of acts.
S ={s1, s2, . . .} is a non-empty set of states.
P = {p1 : A×S→[0,1], p2 : A×S→[0,1], . . .} is a set of probability functions.
U = {u1 : A×S→Re,u2 : A×S→Re, . . .} is a set of utility functions.

An act is, intuitively speaking, an uncertain prospect. I shall use the term ‘uncertain prospect’ when I speak of alternatives in a loose sense. The term ‘act’ is used when more precision is required.

By introducing the quadruple , a substantial assumption is made. The assumption implies that everything that matters in rational decision making can be represented by the four sets A,S,P,U. Hence, nothing except acts, states, and numerical representations of partial beliefs and desires is allowed to be of any relevance. This is a standard assumption in decision theory.

A formal decision problem under risk is a quadruple π =  in which each set P and U has exactly one element. A formal decision problem under uncertainty is a quadruple π = A,S,P,U in which P = /0 and U has exactly one element. Since P and U are sets of functions rather than single functions, agents are allowed to consider several alternative probability and utility measures in a formal decision problem. This set-up can thus model what is sometimes referred to as ‘epistemic risks’, i.e. cases in which there are several alternative probability and utility functions that describe a given situation. Note that the concept of an ‘outcome’ or ‘consequence’ of an act is not explicitly employed in a formal decision problem. Instead, utilities are assigned to ordered pairs of acts and states. Also note that it is not taken for granted that the elements in A and S (which may be either finite or countable infinite sets) must be jointly exhaustive and mutually exclusive. To determine whether such requirements ought to be levied or not is a normative issue that will be analysed in greater detail in subsequent chapters.

It is easy to jump to the conclusion that the framework employed here presupposes that the probability and utility functions have been derived ex ante (i.e. before any preferences over alternatives have been stated), rather than ex post (i.e. after the agent has stated his preferences over the available acts). However, the sets P and U are allowed to be empty at the beginning of a representation process, and can be successively expanded with functions obtained at a later stage.

The formal set-up outlined above is one of many alternatives. In Savage’s theory, the fundamental elements of a formal decision problem are taken to be a set S of states of the world and a set F of consequences. Acts are then defined as functions from S to F. No probability or utility functions are included in the formal decision problem. These are instead derived ‘within’ the theory by using the agent’s preferences over uncertain prospects (that is, risky acts).

Jeffrey’s theory is, in contrast to Savage’s, homogenous in the sense that all elements of a formal decision problem—e.g. acts, outcomes, probabilities, and utilities—are defined on the same set of entities, namely a set of propositions. For instance, ‘[a]n act is . . . a proposition which it is within the agent’s power to make true if he pleases’, and to hold it probable that it will rain tomorrow is ‘to have a particular attitude toward the proposition that it will rain tomorrow’. In line with this, the conjunction B∧C of the propositions B and C is interpreted as the set-theoretic intersection of the possible worlds in which B and C are true, and so on. Note that Jeffrey’s way of conceiving an act implies that all consequences of an act in a decision problem under certainty are acts themselves, since the agent can make those propositions true if he pleases. Therefore, the distinction between acts on the one hand and consequences on the other cannot be upheld in his terminology, which appears to be a drawback.

However, irrespective of this, the homogenous character of Jeffrey’s set-up is no decisive reason for preferring it to Savage’s, since the latter can easily be reconstructed as a homogenous theory by widening the concept of a state. The consequence of having a six-egg omelette can, for example, be conceived as a state in which the agent enjoys a six-egg omelette; acts can then be defined as functions from states to states. A similar manoeuvre can, mutatis mutandis, be carried out for the quadruple <A,S,P,U>.

I think it is more reasonable to take states of the world, rather than propositions, to be the basic building blocks of formal decision problems. States are what ultimately matter for agents, and states are less opaque from a metaphysical point of view. Propositions have no spatio-temporal location, and one cannot get into direct acquaintance with them. Arguably, the main reason for preferring Jeffrey’s set-up would be that things then become more convenient from a technical point of view, since it is easy to perform logical operations on propositions. In my humble opinion, however, technical convenience is not the right kind of reason for making metaphysical choices.

An additional reason for conceiving of a formal decision problem as a quadruple <A,S,P,U>, rather than in the way proposed by Savage or Jeffrey, is that it is neutral with regard to the controversy over causal and evidential decision theory. To take a stand on that issue would be beyond the scope of the present work. In Savage’s theory, which has been claimed to be ‘the leading example of a causal decision theory’, it is explicitly assumed that states are probabilistically independent of acts, since acts are conceived of as functions from states to consequences. This requirement makes sense only if one thinks that agents should take beliefs about causal relations into account: Acts and states are, in this type of theory, two independent entities that together cause outcomes. In an evidential theory such as Jeffrey’s, the probability of a consequence is allowed to be affected by what act is chosen; the agent’s beliefs about causal relations play no role. Evidential and causal decision theories come to different conclusions in Newcomb-style problems. For a realistic example, consider the smoking-caused-by-genetic-defect problem: Suppose that there is some genetic defect that is known to cause both lung cancer and the drive to smoke. In this case, the fact that 80 percent of all smokers suffer from lung cancer should not prevent a causal decision theorist from starting to smoke, since (i) one either has that genetic defect or not, and (ii) there is a small enjoyment associated with smoking, and (iii) the probability of lung cancer is not affected by one’s choice. An evidential decision theorist would, on the contrary, conclude (incorrectly) that if you start to smoke, there is an 80 percent risk that you will contract lung cancer.

It seems obvious that causal decision theory, but not its evidential rival (in its most na¨ıve version), comes to the right conclusion in the smoking-caused-bygenetic-defect problem. Some authors have proposed that for precisely this reason, evidential decision theory should be interpreted as a theory of valuation rather than as a theory of decision. A theory of valuation ranks a set of alternative acts with regard to how good or bad they are in some relevant sense, but it does not prescribe any acts. Valuation should, arguably, be ultimately linked to decision, but there is no conceptual mistake involved in separating the two questions.

A significant advantage of e.g. Jeffrey’s evidential theory, both when interpreted as a theory of decision and as a theory of valuation, is that it does not require that we understand what causality is. The concept of causality plays no role in this theory.

Since there are significant pros and cons for either the causal or evidential approach, it seems reasonable to opt for a set-up that allows for both kinds of theories, until the dispute has been resolved...

Replies from: APMason
comment by APMason · 2012-05-27T01:33:11.360Z · LW(p) · GW(p)

Hmm, if I've understood this correctly, it's the way I've always thought about decision theory for as long as I've had a concept of expected utility maximisation. Which makes me think I must have missed some important aspect of the ex post version.

comment by lukeprog · 2012-05-26T23:41:11.712Z · LW(p) · GW(p)

The objection to the standard approach is put nicely in chapter 10 of Peterson (2009):

Strictly speaking, when Bayesians claim that rational decision makers behave 'as if' they act from a subjective probability function and a utility function, and maximise subjective expected utility, this is merely meant to prove that the agent's preferences over alternative acts can be described by a representation theorem and a corresponding uniqueness theorem... Bayesians do not claim that the probability and utility functions constitute genuine reasons for choosing one alternative over another. This seems to indicate that Bayesian decision theory cannot offer decision makers any genuine action guidance...

Briefly put, the objection holds that Bayesians 'put the card before the horse' from the point of view of the deliberating agent. A decision maker who is able to state a complete preference ordering over uncertain prospects, as required by Bayesian theories, already knows what to do. Therefore, Bayesian decision makers do not get any new, action-guiding information from their theories.

...the output of a decision theory based on the Bayesian approach is a (set of) probability and utility function(s) that can be used to describe [a rational agent] as an expected utility maximizer... What Bayesians use as input data to their theories is exactly what a decision theorist would like to obtain as output.

...a Bayesian decision theory has strictly speaking nothing to tell [an agent] about how to behave. However, despite this, Bayesian frequently argue that are representation theorem and its corresponding uniqueness theorem are normatively relevant for a non-ideal agent in indirect ways. Suppose, for instance, that [an agent] has access to some of his preferences over uncertain prospects, but not all, and also assume that he has partial information about his utility and probability functions. Then, the Bayesian representation theorems can, it is sometimes suggested, but put to work to 'fill the missing gaps' of a preference ordering, utility function and probability function, by using the initially incomplete information to reason back and forth, thereby making the preference ordering and the functions less incomplete. In this process, some preferences for risky acts might be found to be inconsistent with the initial preference ordering, and for this reason be ruled out as illegitimate.

That said, this indirect use of Bayesian decision theory seems to suffer from at least two weaknesses. First, it is perhaps too optimistic to assume that the decision maker's initial information always is sufficiently rich to allow him to fill all the gaps in his preference ordering... The second weakness is that even if the initial information happens to be sufficiently rich to fill the gaps, this manouevre offers no theoretical justification for the initial preference ordering over risky acts. Why should the initial preferences be retained? ...As long as no preference violates the structural constraints stipulated by the Bayesian theory, everything is fine. But is not this view of practical rationality a bit too uninformative?

I don't currently view the two objections in the second-to-last paragraph above as necessitating a non-Bayesian decision theory, but the original concern that Bayesian decision theory doesn't technically offer any action-guidance is serious, and the primary motivation for my interest in Peterson's ex ante approach.

Replies from: cousin_it
comment by cousin_it · 2012-05-27T20:59:09.535Z · LW(p) · GW(p)

It's not obvious to me that the ex ante approach would offer more action-guidance for FAI. Our preferences over acts seem easier to observe than our internal utilities over outcomes. An extrapolation effort might use both kinds of data, of course.

Replies from: lukeprog
comment by lukeprog · 2012-05-28T01:53:33.711Z · LW(p) · GW(p)

For the moment I was just thinking of the ex ante approach in the context of offering action guidance to humans. The ex post approach can't offer any direct advice for what to do because an agent that can state its preferences over acts already knows what to do. What I want to do is state how much I value different outcomes and what probability distributions I have over states of affairs, and have a decision theory tell me which action I can take to maximize my expected utility. It seems that Peterson's ex ante approach is the only approach that can provide this for me.

comment by lukeprog · 2012-05-27T00:08:16.257Z · LW(p) · GW(p)

The final page of Peterson (2009) is also quotable:

Surely people could learn more about decision theory and eventually become more rational! In fact, this textbook is part of an attempt to make that happen.